[WebCodecs] Allow encoder to drop some frames in case of realtime mode

[youennf]

No description provided.

[Djuffin]

I haven't found in the spec how exactly VideoEncoder is supposed to signal that the frame is dropped.
This PR assumes that the output callback is never called for dropped frames.
I think it's better to still call the output callback with empty data to make frame drop more transparent.

Was this previously discussed?

[padenot]

This is correct per spec.

4.1. Attempt to use [[codec implementation]] to encode the frameClone according to options.
[...] error handling -- this is not considered an error
4.5. Let encoded outputs be a list of encoded video data outputs emitted by [[codec implementation]].
4.6. If encoded outputs is not empty, queue a task to run the Output EncodedVideoChunks algorithm with encoded outputs.

Step 4.5, 4.6, the implementation has some output, it might have decided to have dropped some input (this is legal per https://w3c.github.io/webcodecs/#dom-latencymode-realtime), and if it has some output, it calls the callback.

We discussed this, e.g. in your message here: w3c/webcodecs#240 (comment) (and you had the same opinion, but also proposed the current specified behavior).

VideoEncoder doesn't signal that a frame has been dropped, there is simply no output. Authors can match timestamps if it wants to, but that's not very explicit. Happy to discuss something at a spec level.